Company memory for engineering teams
Stop paying for the same problem twice.
Anuvrit remembers every fix, decision and lesson your team has made, with the source attached, and hands it to Claude Code, Codex, Cursor and your developers before they start.
Built for SaaS teams · Agencies · AI coding agents
Example. Asked by Claude Code, via MCP: engineering_preflight("move user sessions from Postgres to Redis")
Answer: This was rejected in May. Redis eviction logged users out under load. If you still proceed, use a separate instance with eviction off. The current version is proven in production and reused 3 times.
Sources: DECISION · MAY, Keep sessions in Postgres; LESSON · REVIEWED, Eviction logged users out; TRUST RECORD, Current code, 3 safe reuses.
Agent checked first, no repeat mistake.
Example. Asked by a developer: Webhook retries are charging some customers twice. Have we seen this before?
Answer: Yes, in March. Same cause: retries sent without an idempotency key. The fix exists. PR #412 stores the key before calling the provider. It was never applied to the new billing-v2 service. That is the gap.
Sources: INCIDENT · MAR 14, Duplicate charges on retry; PR #412 · MERGED, Store idempotency key first; ADR-019 · ACCEPTED, How payment retries work.
Known fix found, nothing re-debugged.
Example. Asked by a tech lead: Starting a Shopify checkout for a new client. What went wrong the last few times?
Answer: Discount sync failed on large carts twice. Batch the calls. Start from the Client F module. It is proven in production. Client code stays walled. Only the reviewed lesson crosses over.
Sources: RETRO · CLIENT D, Discount sync broke at scale; LESSON · REVIEWED, Batch cart calls over 100 lines; MODULE · CLIENT F, Checkout build, in production.
Proven build reused, client data stays private.
Example. Reported by Claude Code, mid-task: Usage limit reached. Task "add audit log export", step 3 of 5.
Answer: Checkpoint saved. Goal, plan, decisions and next step, no transcript. Steps 3 to 5 continue in Codex on the models the plan chose. Nothing is redone. The first two steps stay verified.
Sources: CHECKPOINT, Goal, plan and decisions; STEPS 1-2, Tests and lint passed; NEXT · CODEX, Export endpoint.
Picked up in Codex, nothing re‑explained.
Works inside the agents you already use.
Anuvrit doesn't write code itself. It gives your agents your team's history, runs them on the right model, and keeps work moving when a tool hits its limit.
Anuvrit is not a coding agent. You connect it over MCP to the agents you already use. Before an agent acts, it receives your past decisions, fixes and reviewed lessons with their sources. The Specialised Router plans the work, runs each step on the right model through your own tools, and hands the task to another tool when one hits its limit. Afterwards the outcome is written back.
Memory
Past decisions, fixes and reviewed lessons, with sources, before the agent acts.
Specialised Router
Plans each step, runs it on the right model through your own tools, and keeps the work moving when a tool hits its limit. It does not write code itself.
Two ways to use it. Memory and routing.
Memory gives agents and developers your history before work starts. The Specialised Router runs each step of the work on the right model and keeps it moving across tools. Both read from and write to the same company memory.
Memory
Connect Claude Code, Codex or Cursor over MCP, or ask in plain language. The answer comes from your own sources.
- Before the agent edits
The agent receives the earlier fixes, decisions and reviewed lessons that matter, each with its source, and records the accepted outcome afterwards.
- When a developer asks
“Why did we decide this?” or “Have we fixed this before?” The answer cites the pull request, task or document behind it, and each citation is checked before it is shown.
Routing
The Specialised Router connects to Claude Code, Codex and Cursor, plans the work and runs each step through the tools you already use.
- The right model for each step
A plan with the model and reasoning effort for each step, and a one-line reason. Each step runs on that model, and one task can use several.
- Checked, handed off, finished
Steps are checked with your own tests, lint and type checks. When a tool hits its limit, you resume in another with nothing to re-explain, and every task ends with a cost line.
Built for teams that repeat themselves.
SaaS teams repeat themselves over time. Agencies repeat themselves across clients. The same memory helps both.
The same problems come back over months and years.
The incident that comes back
A timeout you fixed in one service appears again in a new one. The agent sees the earlier fix and its reasoning before it writes the handler.
“Why did we decide this?”
Eight months later, someone asks why the team chose a queue over a scheduled job. The answer arrives with the pull request and discussion behind it.
People move on, decisions stay
When an engineer leaves, their decisions stay searchable with their sources. Agents check that history before changing billing code.
Illustrative examples.
The same builds come back from one client to the next.
The same build, another client
Stripe, Shopify, auth, a CMS. The lessons from the last build are ready before the next one starts.
Client walls stay up
Client A's code is only returned to people and agents with access to Client A, because tenant and scope access is checked on every read.
Joining mid project
A developer who joins halfway through gets the project history in minutes, with the decisions and the reasons for them.
Illustrative examples.
Why you can trust the answers.
An answer is only useful if you can check it. Every one is built from your own sources, and the evidence comes with it.
Every claim is cited
Each answer points to the exact source and version it came from, and every citation is checked before it is shown.
Unsupported claims are dropped
If the sources do not support a statement, it is left out and the gap is named, instead of being filled with a guess.
Access checked on every read
Tenant and scope access is checked each time memory is read, so people and agents only see what they are allowed to.
Trust records for reused code
Shows where reused code came from and why it is trusted.
What has shipped. What is ahead.
Anuvrit is in beta. These are the milestones so far, and the short list of what we are working on next.
Shipped
- Capture and retrieval
- Engineering memory for agents
- Production release and public site
- Lossless, versioned archive
- Histories and trusted reuse
What is ahead
- Self-serve sign-up and accounts
- Connecting your tools yourself, with progress as they are read
- A hosted MCP endpoint and a setup guide for each coding agent
- A graph view and a dashboard of what your team has learned
- Plans, billing and the Specialised Router add-on at public launch
Give your agents your team’s history.
Anuvrit is in beta and opening to engineering teams in stages. Join the waitlist and we will contact you when there is a place for your team.